Can probability describe a fact that already happened?

Yes. It can express an analyst's uncertainty about an unknown past fact.

Must every Bayesian use the same prior?

No. Different information or prior constructions can produce different assignments, which should be made explicit.

Does more data always remove prior differences?

Not necessarily. Identifiability, model assumptions, prior support, and the information in the data all matter.

AI Glossary / Statistics and probability / Intermediate

Bayesian probability

Definition

Bayesian probability interprets probability as a quantified degree of uncertainty about a proposition, conditional on available information and modeling assumptions. It can describe uncertainty about a future event, an unknown parameter, or a fact that is already settled but not known to the analyst. Probabilities must obey consistent mathematical rules and can be updated when evidence becomes available.

What it means

A statement such as “the shipment has a 30% probability of arriving late” expresses uncertainty given what is currently known. It does not mean that this particular shipment will be partly late. Once its arrival status is established, the proposition has a definite answer.

Bayesian probability provides a language for connecting such judgments with evidence. Relevant information might include past delivery records, current conditions, and a model of how those conditions relate to delays.

Calling probability subjective does not make every assignment equally well supported. Analysts should explain their information and assumptions. Bayesian approaches differ in how they justify priors, including personal judgments, structured expert elicitation, and reference constructions. The interpretation is broader than any particular statistical software or estimation algorithm.

How it works

Define the proposition precisely, including its time horizon and the information already incorporated. Assign probabilities that respect the probability rules: mutually exclusive exhaustive outcomes have probabilities summing to one, and joint and conditional probabilities must agree.

To update a hypothesis H after observing E, use P(H given E) = P(E given H)P(H)/P(E), provided the denominator is positive. The prior describes uncertainty before this evidence; the likelihood connects possible hypotheses to the evidence; normalization produces the posterior.

Additional observations must be handled using their appropriate conditional relationships. Multiplying likelihoods as if observations were independent can count shared information repeatedly. A posterior can become the prior for a subsequent update when the full information structure is respected.

Review the resulting conclusions against plausible alternative assumptions and, where outcomes can be observed repeatedly, examine predictive performance. Mathematical coherence prevents certain contradictions, but it does not establish that the proposed model captures reality or that its numerical probabilities are well calibrated.

Three practical examples

Project planning: express uncertainty that a milestone will be missed, then revise it when a dependency is completed or delayed.
Equipment monitoring: update uncertainty about an unobserved component condition after receiving a sensor signal.
Forecasting: maintain a distribution over possible demand levels instead of turning an uncertain estimate into a single supposedly certain number.

Think of it this way

Think of probability as a carefully maintained information ledger. Each entry records uncertainty given the information currently available. New evidence changes the balances according to explicit rules. A neatly balanced ledger can still rest on poor inputs, so both the arithmetic and the underlying information need scrutiny.

Worked example

Let H mean that a shipment will arrive late. Before a new alert, an analyst assigns P(H) = 0.30. Suppose the alert E occurs with probability 0.80 for late shipments and 0.20 for shipments that arrive on time.

The joint probability of lateness and an alert is 0.30 × 0.80 = 0.24. The joint probability of being on time and receiving an alert is 0.70 × 0.20 = 0.14. Therefore P(E) = 0.38, and P(H given E) = 0.24/0.38, approximately 0.631579.

The alert increases the assessed probability of lateness from 30% to about 63.16%. It does not make lateness certain. Prior odds of 3:7 are multiplied by the likelihood ratio 0.80/0.20 = 4, producing posterior odds of 12:7, which gives the same answer.

For a sensitivity comparison, keep the alert model fixed but use prior probability 0.60. The posterior becomes 0.48/(0.48 + 0.08), approximately 0.857143. This difference exposes the influence of the prior information. It is not a mathematical inconsistency, and the two prior assessments still need substantive justification.

Illustrative case study

Fictional case study: Harbor Parcel asks planners to record shipment-delay probabilities before a new alert system is introduced. They define “late” using an agreed delivery deadline and document which route information is already included.

During the pilot, the team compares predictions with completed deliveries and checks performance separately across routes. It discovers that one alert partly repeats information already used in the initial assessment. The model is revised to represent that dependence instead of treating the alert as entirely fresh evidence.

The exercise illustrates how explicit probability judgments can reveal hidden assumptions. It does not claim that a particular posterior guarantees a good operational decision.

Common misconceptions and mistakes

Do not confuse a probability with a degree of truth. A shipment is not 63% late merely because its lateness probability is 0.63.

Do not assume that applying Bayes' theorem alone distinguishes Bayesian from all other statistical reasoning; the theorem is a general probability identity. The interpretation and modeling treatment of uncertainty matter.

Avoid equating coherence with empirical accuracy, treating expert confidence as evidence without examination, or multiplying dependent reports as independent confirmations. A posterior probability is conditional on the specified model and information, not an unconditional certificate of truth.

Business use

Business teams can use Bayesian probability to communicate uncertainty, compare assumptions, and update forecasts consistently. The practical benefit comes from making the proposition, evidence, and reasoning visible enough to challenge.

Decisions also require consequences and preferences. A 30% disruption probability could justify action when prevention is inexpensive and disruption is costly, while a higher probability might not justify an expensive intervention with little benefit. Probability informs that comparison; it does not determine the action by itself.

Check your understanding

What changes the prior odds in the worked example?

Answer: The alert's likelihood ratio of four.

Why do the two posterior probabilities differ?

Answer: They start from different prior probabilities while using the same alert model.

Does a coherent probability automatically select an action?

Answer: No. Decision consequences and objectives are also needed.

Related terms

Learn next

distinguish uncertainty about parameters from variability in observations, practice odds-form updating, and examine sensitivity to priors and dependence assumptions. Then connect probabilistic forecasts to explicit decision costs.

Sources

Sources: David Poole and Alan Mackworth, Artificial Intelligence: Foundations of Computational Agents, third edition, Section 9.1, https://www.cs.ubc.ca/~poole/aibook/3e/html/ArtInt3e.Ch9.S1.html . The author-hosted account of Bayesian probability, probability rules, and conditional updating was read. The shipment example and prior-sensitivity calculations are original and independently checked.

Previous: Bayesian optimization acquisition function Next: Bayesian program synthesis